FIT-IRSA: Feedback-Integrated Two-Phase IRSA with Deep Reinforcement Learning

Efficient random access can be used in scenarios with a massive number of IoT devices. Among modern random access protocols, Irregular Repetition Slotted ALOHA (IRSA) offers excellent asymptotic performance (for large frame sizes), but its finite-frame efficiency is lower and difficult to optimize analytically. In this work, we introduce limited mid-frame feedback to better coordinate users and improve performance: Feedback-Integrated Two-phase IRSA (FIT-IRSA). We formulate IRSA with feedback as a deep reinforcement learning (DRL) problem. Using policy gradient methods, we learn transmission strategies that improve throughput under varying loads, as demonstrated in our simulation results. This provides a practical alternative to classical density-evolution–based optimization, which applies mainly to large frames.